jammi-ai 0.34.0

The Python client for a Jammi engine (import `jammi`) — one connect() for local (embedded) and remote targets

jammi-ai

The Python client (import jammi) for a Jammi engine — one connect(target) for local (embedded) and remote targets.

jammi-ai is the base client: a universal py3-none-any wheel that links no candle/ML stack — just grpcio + protobuf + pyarrow — for production where the engine runs behind a server. It also discovers the compiled in-process engine as an OPTIONAL backend: pip install jammi-ai[embedded] pulls jammi-ai-native, and then connect("file://…") resolves the local target to an in-process engine (direct FFI) through the SAME front door. Absent the extra, the base stays native-free and a file:// target raises a truthful NoEmbeddedEngineError.

One front door

import jammi

db = jammi.connect("https://engine.example.com")
db.add_source("patents", url="s3://corpus/patents.parquet", format="parquet")
db.generate_embeddings(source="patents", model="local:tiny_bert",
                       columns=["abstract"], key="id", modality="text")
q = db.encode_query(model="local:tiny_bert", query="quantum computing")
hits = db.search("patents", query=q, k=5)   # -> pyarrow.Table

connect(target) is the Python mirror of the Rust Jammi::open(Target): transport is config, not a code path.

Authenticated channels

A bearer-protected endpoint is reached by attaching credentials to the channel — the bearer rides the channel (attached once at connect), not threaded through every call:

from jammi import connect, BearerCredentials

db = connect("https://engine.example.com", credentials=BearerCredentials(token))

The same works on a plaintext grpc:// target for local development. The channel-level bearer covers the typed gRPC verbs; db.sql() (the Flight SQL lane) does not yet carry it — tracked at issue #96.

target	transport
https://host / grpcs://host:8081	secure remote
http://host / grpc://host:8081	plaintext remote
file:///data	in-process engine (direct FFI) with the [embedded] extra installed; otherwise raises NoEmbeddedEngineError pointing at pip install jammi-ai[embedded]

Scaling local→remote is an env flip (connect(os.environ["JAMMI_TARGET"])) with no code change. The lean deploy build and the embedded build are ONE package — the same import jammi, connect unchanged — differing only by whether the [embedded] extra is present.

Generated from the canonical proto

The wire stubs under jammi/_generated/ are generated from the same jammi.v1 proto the engine, npm client, and embed wheel speak — one source, no parallel schema. Regenerate after a proto change:

pip install -e '.[dev]'
make generate